#	Pagina
attuale pagina	/open-h2020/projects/199451/results.html

Report

Teaser, summary, work performed and final results

Periodic Reporting for period 2 - NASDAC (iNnovative Approaches for Scalable Data Assimilation in oCeanography)

Teaser

Summary

The project focus is on the Predictive Science as the paradigm shift of the emerging CSE (Computational Science and Engineering) which tightly integrates the numerical simulations of Computational Science and Engineering with Validation and Verification and Uncertainty Quantication (UQ). UQ is an essential ingredient of Predictive Science, whose aim is to not only reproduce with high fidelity an observed phenomenon, but also to predict the reality in absence of measurements. To this end, reliable numerical predictions require complex nonlinear physical models as well as a systematic and comprehensive treatment of calibration and validation procedures, including the quantication of inherent uncertainties. Furthermore, because the equations governing physical model contain multiscale, multilevel nonlinear spatio-temporal interactions that use ever more data and ner model grid resolutions, due to the inherent ill conditioning of the underlying mathematical models, increasing size of the computational data leads to increasing amount of uncertainties. Therefore, it is crucial to assess the impact of these uncertainties on future predictions.

The planned result was the redesign of the software stack, the main innovations being situated at the medium and low level of the stack, ranging from the simultaneous introduction of space-and-time decomposition approaches (i.e. parallel in time (PINT) methods coupled with Hybrid Data Assimilation models (Ensemble and Variational), composition (additive or multiplicative) of Block Communication Avoiding Algorithms for preconditioned high-order nonlinear solvers that perform more computation to obtain greater accuracy for each computational degree of freedom; and additionally, at the lowest level, the reuse of recent scalable linear algebra solvers developed by other EU-funded and still active projects. The planned results aimed at changing the way we think about the computational approach to simulation problems. Rather than applying more resources to an existing formulation to obtain a more accurate solution or to solve a larger problem, the proposed activity has provided an opportunity to loosen the grip of, or even remove, computationally-imposed simplications.

Work performed

The focus has been on scalable algorithms that exploit space-and-time parallelism (i.e. the so-called Parallel In Time - PINT, approaches) for performing Predictive Science on extreme scale HPC systems. Because data movement consumes much more energy than arithmetic operations and memory accesses are increasingly the bottleneck in computational algorithms, the performed work also addressed cross cutting issues concerning hierarchical algorithms which reduce communication and task synchronization to maximize the number of useful calculations per memory accesses.

Note that because of crucial changes in the beneficiaries\' staff, with key personnel unexpectedly moving to other organizations in the middle of the project, the initial secondment plan (and the related management effort) could not be properly sustained. This justifies the deviation from the original workplan, the missing of milestones and deliverable submission, as listed in the initial description of work, and eventually the early termination of the project. This particularly applies to D2, D6-D8, D10-D15 as well as M2, M4-M6, M8, M10-M11.

Final results

The Software Stack for preparing Applications towards Exascale Computing focusing on the overall description of the software stack for transition toward exascale computing. The software stack represents the multilevel architecture for transitions of applications to exascale. The research activity addressed the design and development of innovative approaches across the overall software stack. Innovations wee situated at the medium and low level of the stack, ranging from the simultaneous introduction of space-and-time decomposition approaches (i.e. parallel in time (PINT) methods coupled with Hybrid Data Assimilation models (Ensemble and Variational), composition (additive or multiplicative) of Block Communication Avoiding Algorithms for pre-conditioned nonlinear solvers, and finally, at the lowest level, the reuse of recent scalable linear algebra solvers developed by other EU-funded and still active projects.